Individual-Referential Evaluation Protocol · open standard · v0.1
I represent myself.
name · photo · age · sex · nationality · address · highest degree · what I have actually done, and how I got here
IREP is an open protocol for hiring and admissions. It evaluates the person — evidence, work, trajectory — while category labels stay out of evaluation, inside a separate firewalled audit channel, unavailable until the decision is final.
EU AI Act · Annex III, point 4
AI systems used to recruit, screen applications or evaluate candidates are listed among the Annex III high-risk use cases. Under the 2026 Digital Omnibus, the corresponding Chapter III obligations apply from 2 December 2027. IREP's rule does not wait for that date: AI may assist; humans decide.
Official sources: AI Act, consolidated text · European Commission implementation timeline. Whether a given system is high-risk also depends on Article 6(3); implementers remain responsible for their own assessment.
Thirty years of unread memos
The research existed. The sector never applied it.
Merely recording your race on a form before a test measurably depressed performance. A checkbox suffices to activate a stereotype.
Steele & Aronson, 1995 · JPSPIdentical CVs, only the name changed: callback rates diverge sharply. Twenty-five years of audit studies show the gap essentially unchanged.
Bertrand & Mullainathan, 2004 · Quillian et al., 2017France tested anonymous CVs at national scale. Outcomes for immigrant-background candidates got worse — removing labels alone was not a sufficient remedy, and it can remove helpful context too.
Behaghel, Crépon & Le Barbanchon, 2015The problem was never seeing or not seeing the category. It is the moment we let it speak. IREP adopts the strict separation algorithmic fairness converged toward: category data may support lawful, firewalled audits — never evaluation-stage decisions.
The protocol
Evaluation first. Audit after decision. Firewalled.
Category-blind, trajectory-rich screening
No name, photo, age, sex, nationality, or credential labels. Full richness of what the person has done — gaps, pivots, self-taught depth included.
Untimed structured assessment
Bounded window, no per-item pressure. Validated, pre-registered rubrics only. Time pressure measures stress, not competence.
Decision on evidence
Documented rationale, by evaluators with zero access to category data. Humans decide.
Firewalled measurement
Category data flows through a separate, firewalled audit channel — voluntary, lawful, analytically unavailable until decisions are final. Contracts and equity audits only.
separate audit channel — unavailable until the decision is final
The individual is the unit of interpretation. The role is the standard of decision. The population is the instrument of audit. And the question is not only “compared to what?” — it is “compared to when?”
For recruiters & talent teams
You could be running this by tomorrow.
No new software, no bigger budget, no permission needed. Four moves on your next role — then climb: Starter (~2 weeks), Aligned (~2 months), Steward (~6 months, audits published). All self-declared, like a license: nobody certifies you, and you answer publicly for it.
Offered to the commons
IREP is not a product. It is not a company. It is not an asset to be owned. It is intentionally offered to the commons.
Free, open source, permanently — including for commercial HR vendors building it into their own tools. Behind every hiring decision there are one or several lives engaged. Access to employment, like access to education, is a gateway a society cannot equitably price.
The initiator is the initial steward — not the owner: to document it, test it, and convene the people willing to improve it — before handing stewardship to a non-profit collective.
The honest part. The repo ships a synthetic demonstration — labeled loudly as a demonstration, never a validation: the biases are injected, so detecting them is circular. The premise now has a first real-data measurement, Stage 0.5 of the validation plan: on 69 seasons of public baseball data, a pilot measured where an individual's own record overtakes their reference class out of sample. Hypothesis-generating, one domain; the confirmatory test is a second domain with a different noise structure. Real evidence follows the validation plan: retrospective re-scoring, consented parallel pilots, a pre-registered RCT. Near-term claim: IREP selects differently, traceably. “Better” takes longitudinal data — the protocol is judged on its trajectory, as it preaches. Null results get published.
In the field
Independent perspectives.
Third-party writing about IREP is listed as received. It is commentary, never validation evidence, and listing is not an endorsement in either direction.
“The Monoculture Tax” · Tyler Style, solutions architect and engineering leader, Vancouver, WA. LinkedIn, 7 August 2026. A practitioner essay on how workplace culture drifts toward monoculture through hiring and promotion. He reviewed the IREP instrument at the initiator's request, notes that it is early (v0.1, not yet validated), and writes that it “evaluates each person against their own evidence and trajectory”. He then sketches carrying IREP's rubric structure into promotion, an extension beyond v0.1's stated scope of hiring and admissions.
Three circles
It needs people more than it needs code.
Practitioners
Recruiters and headhunters who never trusted the tooling. You are not ambassadors — you are the pilot sites where the head-to-head validation happens.
Open a pilot-site issue →Researchers
Labor economics, I/O psychology, psychometrics, ML fairness. The field-experiment infrastructure is pre-built; the analysis plans want pre-registering.
Open a research-partner issue →Voices
People who write about hiring, credentialism and the commons. Read the spec. Criticize it in public — sharp issues are the contribution.
Read & critique →Not sure which circle is yours? Start where everyone starts: introduce yourself, ask a question, or challenge the spec — in the open.
Join the discussion →